The UK Court of Appeal has convened to hear arguments in the landmark appeal of Getty Images v Stability AI, a defining legal showdown that will establish the global common-law precedent on whether artificial intelligence model weights constitute an “infringing copy” under English copyright law.
The central legal question: Are neural weights infringing copies?
Following Mrs. Justice Joanna Smith’s High Court ruling in the Chancery Division, the appeal centers strictly upon an unprecedented statutory interpretation of the Copyright, Designs and Patents Act 1988 (CDPA):
- Interpretation of Section 27 CDPA: The appellate court must decide whether mathematical parameters (weights and biases) that capture the latent artistic features of millions of copyrighted images qualify as an “infringing copy” when imported or distributed in the UK, even though the underlying raw pixel files are not stored within the model.
- Latent representations vs. storage: Stability AI argues that diffusion model weights are statistical abstractions that do not contain or reproduce original works. Getty Images contends that compiling expressive visual relationships into commercial model parameters without a license constitutes unlawful secondary infringement upon distribution.
- Implications for AI training in the UK: If the Court of Appeal accepts Getty’s statutory construction, commercial AI developers deploying models in the UK will be strictly required to secure licensed consent from copyright holders, fundamentally altering the economics of frontier model deployment.
Practical consequences for technology developers and creators
The Court of Appeal’s impending determination carries profound ramifications across the UK creative and technology sectors:
- Licensing frameworks for generative models: A decision favouring content holders will immediately accelerate collective licensing frameworks and automated copyright attribution systems across the UK.
- Supply chain and model importing risk: UK businesses importing or fine-tuning foreign foundation models could face secondary copyright liability if the underlying training corpus included unlicensed proprietary works.
- Judicial scrutiny over algorithmic transparency: Both parties presented detailed technical expert evidence, underscoring that future intellectual property disputes will demand rigorous mathematical documentation of training pipelines and weight extraction methods.